Machine Learning-Based Approach for Rapid Prediction and Analysis of Gas-Sensing Properties of Janus Materials

The environmental and health risks associated with toxic gas emissions highlight the urgent need for high-performance and cost-effective gas-sensing materials. Here, we establish an integrated density functional theory-machine learning (DFT-ML) framework for the accelerated screening of Janus transition metal dichalcogenide materials toward NO, NO₂, and NH₃ sensing. A total of 162 Janus material-gas adsorption systems were established, and adsorption energies were obtained from DFT calculations on representative structures. An initial feature set was derived from intrinsic material properties and gas descriptors. Redundancy was reduced through Pearson correlation analysis, resulting in 25 relevant features for subsequent modeling. Among 12 ML models evaluated, Gradient Boosting Regression (GB) achieved the best predictive performance (<i>R</i>² = 0.960, RMSE = 0.057, MAE = 0.042 on the validation set). Subsequent SHAP analysis revealed that the material band gap (sb) strongly interacted with gas type (gt), average atomic radius (gr), and electronegativity (gn), which were critical for accurate model predictions. With the optimized GB model, promising candidates were rapidly screened and further assessed with respect to adsorption distance, charge transfer, work function, and recovery time. In particular, NbSSe and VSSe were identified as promising materials with excellent responsiveness and reusability for NO and NO₂ detection. These findings underscore the effectiveness of the proposed DFT-ML framework for high-throughput gas-sensing material discovery. In conclusion, this study establishes an effective DFT-ML paradigm for high-throughput discovery of gas-sensing materials, achieving significant reductions in computational cost without sacrificing predictive reliability, and offering a viable pathway toward the intelligent design of complex material systems.